China has a unique national condition: data is in the east; energy is in the west.

Beijing, Shanghai, Shenzhen, and Hangzhou account for over 65% of national AI compute demand, but land, electricity, and cooling costs in these places remain high; while Inner Mongolia, Ningxia, Guizhou, and Gansu have abundant wind, solar, and hydro power along with cool climates, built-out compute clusters sit underutilized -- in 2024, average utilization at western compute hubs was only 30%, with massive GPU idle time in datacenters.

In 2026, this awkward situation is being broken.

East Data West Computing: From Laying Pipelines to Flowing Water

The "East Data West Computing" project launched in 2022; the past four years were mainly "building pipelines" -- laying high-speed fiber networks, constructing eight major compute hubs, connecting ten clusters. But 2026 is the inflection point: the project officially enters scaled operations phase, with western compute hub utilization climbing from 30% to 65%, more than doubling.

The key drivers are the two cloud giants' "nationwide compute scheduling" services. Alibaba Cloud launched "Compute South-North Water Diversion" (internal codename) in June, using a self-developed distributed scheduling engine to intelligently route eastern training and inference jobs to the most cost-effective western nodes; Tencent Cloud followed with "One Compute Network," claiming cross-region scheduling latency reduced by 40% and bandwidth costs down 55%.

What does this mean? An AI startup in Hangzhou doesn't need to rent a datacenter in Guizhou or build its own dedicated line; just click "intelligent scheduling" in the Alibaba Cloud console, and training jobs are automatically assigned to nodes with the lowest electricity prices and most idle GPUs -- just like you don't need to know whether your electricity comes from the Three Gorges or wind power.

Domestic Chips: From Usable to Useful

Another underlying variable in the compute network is domestic chips.

Huawei Ascend 910C completed large-scale deployment in H1 2026. According to public data, Ascend 910C achieves 320 TFLOPS at FP16 precision, a 60% improvement over 910B; memory bandwidth raised to 1.6 TB/s. More critically, the software stack: the CANN operator library now covers over 90% of mainstream AI framework operators, MindSpore and PyTorch adaptation maturity has improved substantially, and developer migration costs are significantly lower.

In inference scenarios, domestic chips have achieved 35% substitution rate -- this isn't a planned figure; it's real operational data. Ascend and Cambricon chips can stably handle standardized inference tasks like recommendation systems, content moderation, speech recognition, and OCR. Training scenario substitution is still around 10%, but in sensitive industries like finance and government, domestic training clusters have begun shouldering core tasks.

This isn't an aggressive number, but it means domestic chips have crossed the "usable" threshold into the "useful" phase. The remaining gap will be closed through scale effects and iteration speed.

Green Compute and Night Discounts

Cost is the sharpest weapon.

Compute hubs in Inner Mongolia and Ningxia have achieved 100% green power supply -- wind, solar, and hydro feeding directly into data centers, with PUE (Power Usage Effectiveness) dropping below 1.1, far exceeding the national average of 1.3. Green power isn't just a carbon neutrality story; it directly means lower electricity prices: western green-powered data center electricity can be as low as 0.28 RMB/kWh, while eastern coastal industrial electricity generally runs 0.7-0.9 RMB/kWh.

Even more aggressive is the "night compute 80% off." To absorb idle overnight wind and solar output (midnight to 6 AM is the lowest grid load period, with large amounts of new energy power wasted), Alibaba Cloud and Tencent Cloud successively launched "Nighttime Spot Instances" priced at just 20% of daytime on-demand instances. A full fine-tune of a 7B model that costs roughly 2,000 RMB running on A100 during the day costs only 200-300 RMB running on Ascend 910C at night.

This echoes the LLM pricing polarization trend (see this issue's Brief "LLM Pricing Polarization"): when Doubao and others slash API prices to commodity levels, the foundation is collapsing compute costs. For SMEs and independent developers, this is the best of times -- training a custom small model costs less than hiring an intern for a month.

New Moats After Compute Equality

But one easily overlooked issue: cheap compute doesn't mean AI becomes easy.

When compute is as on-demand and cheap as water and electricity, "being able to afford model training" ceases to be a competitive moat. Then what is the moat? Application scenarios, data, and industry understanding.

You can fine-tune a medical Q&A model for 200 RMB, but do you have high-quality medical labeled data? Can you get real hospital consultation records? Do you understand clinical pathways and medical insurance rules? These are things compute can't solve.

This is also why Alibaba Cloud and Tencent Cloud, while doing compute scheduling, are furiously stacking industry solutions on top -- they know full well that selling compute is a pipeline business with razor-thin margins; selling scenario solutions is where real business lives. The U.S. market is already validating this logic: AWS gross margins are around 30%, while vertical SaaS like Veeva has gross margins above 70%.

2026 is the inflection point for China's compute infrastructure. East Data West Computing moves from construction to operation; domestic chips move from pilot to scale; green compute moves from PPT to electricity bills. But for AI founders, the real lesson is: don't build a moat on compute, because something becoming as cheap as air can't be defended.

See you tomorrow.